Where Human Judgement Still Matters

Automation debates tend to collapse into a single axis — can a task be automated, yes or no — when the more useful question splits into two separate ones: can it be automated, and should it be. Those two questions get conflated constantly, and the gap between them is where a lot of bad automation decisions get made.

Fact: the tasks most resistant to automation share common features — they involve genuine trade-offs between competing values rather than a single optimisable metric, they carry consequences that fall on people who didn't get a say in the decision, or they require context that exists in relationships and history rather than in any dataset. These aren't tasks that current AI simply hasn't gotten good enough at yet; they're tasks where "good enough" isn't well-defined in the first place, because the right answer depends on whose interests are being weighed and how.

The should question is separate from the can question

Even where a task technically could be automated — a decision that affects someone's livelihood, a judgement call about how to treat a person in a difficult situation, a choice with genuine ethical weight — there's a separate argument for keeping a human accountable for it regardless of whether an algorithm could produce a statistically similar outcome. Accountability itself is part of what the task requires: someone who can be asked why, who can be held responsible, and who can exercise discretion when the general rule produces an unfair result in a specific case.

Analysis: the areas where human judgement holds up longest tend to be exactly the areas that were never really mechanical in the first place, even when they looked that way from the outside — a manager deciding how to handle an underperforming team member isn't really doing data analysis, even if data informs the decision; a professional advising a client through a difficult choice isn't really executing a rules engine, even when rules are part of the picture. Automation keeps eating the mechanical layer underneath these roles, which raises rather than lowers the value of the judgement sitting on top of it.

Opinion: the more interesting failure mode isn't automating something that should stay human — it's the subtler version, where the mechanical parts of a job get automated away and the humans left doing it are pushed to imitate a rules engine anyway, because that's what's easy to measure and manage. Preserving genuine judgement requires deliberately protecting the time and authority for it, not just leaving the judgement-heavy parts of a role technically unautomated.

Prediction, held loosely: as AI keeps closing the gap on tasks that look judgement-heavy but are actually pattern-matching in disguise, the genuinely irreducible cases — the ones involving real trade-offs, accountability and context that isn't written down anywhere — will become more visible for what they are, rather than fewer in number. The interesting shift over the next decade probably isn't AI replacing judgement, it's a clearer, harder-to-avoid distinction between decisions that were always mechanical wearing a judgement costume, and decisions that never were.